Figure 2: A pair of IR raw images.
..... SS
I
ume ee M |
Figure 3: Detection of changes in the orthogonal motion
field.
a
EIGEN
Figure 4: The innovation signal of the Kalman filter cor-
responding to the dotted line shown in the previous figure.
One may notice that the four detected changes correspond
to the fronteers of the car wheels.
306
car wheels.
5 Conclusion
In the introduction we have shown that the explicit infor-
mation is characterized by its local and incomplete nature.
The proposed regularization method allows one to complete
the explicit information where it is available.
In the case of images of outdoor scenes, which are often
well textured, the explicit motion information is available
everywhere in the images. A problem arises when process-
ing images of indoor scenes which are very structured and
where texture is very often not available. Let us consider,
for example, the case of an object moving over an homoge-
neous background. The motion is only effectively perceived
on the object boundaries. The information provided by the
boundaries must be propagated inside the object. There-
fore, it is no more an information processing issue - as the
regularization is - but a topological identification problem
aiming at the determination of structural areas in the image.
The next step of our study will put the emphasis on how
to take into account the structural content of an image in
the context of motion estimation.
Acknowledgments
We gratefully acknowledge the kindness of the SAT company
(Société Anonyme de Télécommunications, 41 rue Canta-
grel, F-75013 Paris Cedez) for providing us with sequences
of IR images, some of which are presented in this article.
References
[1] J. Stuller, G. Krishnamurty, “Kalman Filter Formula-
tion of Low-Level Television Image Motion Estimation",
Computer Vision, Graphics and Image Processing, Vol.
21, No. 2, pp 169-204.
[2] M. Basseville, “Detecting Changes in Signals and Sys-
tems - A Survey”, Automatica, Vol. 24, No. 3, pp 309-
326, 1988.
[3] A. N. Netravali, J. D. Robins, “Motion-Compensated
Television Coding: Part 1”, The Bell System Technical
Journal, March 1979.
[4] H. H. Nagel, “On The Estimation of Optical Flow: Re-
lations between Different Approaches and some New Re-
sults”, Artificial Intelligence, 33, pp. 299-324, 1987.
[5] E. C. Hildreth, “Computations Underlying the Measure-
ment of Visual Motion”, Artificial Intelligence, 23, pp.
309-354, 1984.
Abs
for 1
mat
relat
mat
Th
XW
mod
use
Digi
with
Key
com
visic
visi
use
This
digi
stere
degr
et al
imag
clos:
metr
accic
be 2
inter
limit
matc
visio
corre
Thes
softw
1.1
visio
follo
is re
meth
line-
descr
the p:
funct
edge:
edges
